Security & confidentiality
Useful AI in healthcare is AI that puts no one at risk: not the patient, not the institution, not the client company. This page states what our tools do with data, and what they never do.
Disclaimer. This page is not legal advice. Every project goes through the data protection officer (DPO) of the institution or client, who alone decides what is allowed there.
The principle
Identifying patient data never goes into a consumer tool. It stays in the institution’s information system, under the institution’s responsibility.
Our tools work on anonymised, aggregated or non-identifying data: department organisation, teaching, literature watch, manuscripts, protocols. When a use requires more, it moves up a level, with different conditions.
Three usage levels
| Level | Data | Examples | Conditions |
|---|---|---|---|
| Green | No personal data | Organisation, teaching, literature watch, manuscripts | The method’s guardrails are enough |
| Orange | Pseudonymised data | Studies and registries | Contract and DPO opinion; a business plan with contractual data commitments, or a locally installed tool |
| Red | Identifying patient data | Medical record, named reports | Only within the institution’s information system or with a certified health-data host, under contract |
Each service offer shows its level. None is at the red level: that data does not leave the institution.
Which tool for which level
| Tool | Training on your data | Retention | HIPAA contract (BAA) | Data in Europe | HDS certification |
|---|---|---|---|---|---|
| Claude Free, Pro, Max | Yes, unless you opt out in the settings | 5 years if training is accepted, 30 days otherwise | No | No | None found |
| Claude Team | No (commercial terms) | Per contract | No | No | None found |
| Claude Enterprise | No | Per contract | Yes, with the HIPAA configuration: Claude Code in the terminal and Cowork running locally are covered; editor extensions and cloud sessions are not covered | No | None found |
| Claude via AWS Bedrock (EU profile) | No | Per the AWS contract | Via AWS | Yes, routing between EU regions | AWS is HDS-certified; inclusion of Bedrock to be confirmed in writing |
| Claude via Google Vertex AI (EU multi-region) | No | Per the Google contract | Via Google | Yes, EU member states | Google is HDS v2-certified, Vertex AI within the announced scope |
In practice: green with a business account (Team or Enterprise); orange with Enterprise or a European route, under a data processing agreement and an impact assessment approved by the DPO; red only within the institution’s information system or with an HDS-certified host, under the institution’s contract. Consumer accounts must not be used for any sensitive professional data. HIPAA is a US framework: it replaces neither the GDPR, nor HDS, nor medical confidentiality.
Sources, consulted on 7 November 2026: Anthropic consumer terms, HIPAA-ready Enterprise plans, Vertex AI data residency, HDS at Google Cloud, HDS at AWS. Offers change: check before each project.
What we do not do
- No consumer account for sensitive professional data.
- No hosting or retention of health data: we do not operate at the red level.
- No clinical decision support tool: that would be a medical device, with its own obligations.
- No agent that acts without human approval.
The reference framework
| Framework | What it requires, in short | Date | Official source |
|---|---|---|---|
| GDPR and CNIL (French data protection authority) | Health data are sensitive data with reinforced protection; some processing requires prior formalities. CNIL practical guidance on developing and evaluating AI systems that use health data | 5 March 2026 | cnil.fr |
| HAS-CNIL guide “AI in the care setting” | Draft recommendations for institutions and professionals, from choosing a system to decommissioning it, with a governance sheet and a generative-AI sheet | Public consultation closed on 16 April 2026; final version pending | cnil.fr |
| HDS (French health-data hosting certification) | A provider hosting health data on behalf of a third party must be certified; version 2 of the standard notably requires storage within the European Economic Area | End of transition to v2: 16 May 2026 | esante.gouv.fr |
| Decree No. 2026-209 of 24 March 2026 | Hosted health data stored exclusively in the EU or EEA; remote access from a third country is treated as a transfer within the meaning of the GDPR; expanded contractual clauses; electronic archiving included | Archiving: 27 March 2026; EU/EEA storage: 26 September 2026 | legifrance.gouv.fr |
| EU AI Act | Transparency obligations (Article 50) apply; the high-risk regime was postponed by Regulation (EU) 2026/1744, the “Omnibus” | Article 50: 2 August 2026; high risk: 2 December 2027 (standalone systems) and 2 August 2028 (regulated products, including medical devices) | eur-lex.europa.eu · Commission timeline |
| HIPAA (US clients) | A Business Associate Agreement is required with any provider, including AI or cloud vendors, that handles protected health information | In force | hhs.gov |
| OWASP Top 10 for Agentic Applications | The ten risks specific to AI agents (goal hijacking, tool misuse, excessive privileges…) | Published in December 2025 | genai.owasp.org |
Page updated on 7 November 2026. Texts change: check the version in force on the official source before relying on it.
Agent security
An agent that reads your emails and files can be hijacked by what it reads. Five rules, applied to every tool:
- Least agency. The agent has only the rights its task needs; every unnecessary capability is an attack surface.
- External content is data, never an instruction. What an email, a web page or a document asks for is not executed: it is something to analyse.
- Human approval before any send or write. The agent prepares, a person accepts.
- An action log. What was done, where, and how it was checked, in four lines at the end of each session.
- No file is ever overwritten. A corrected version is a new file; nothing is deleted, everything is archived.
For companies
A health company or start-up adds client confidentiality to patient confidentiality:
- Client confidentiality. No client, product or project name in a consumer tool, a shared document or training material.
- Pseudonymised data. They stay at the orange level: contract, DPO, suitable tool.
- Traceability. Every opinion is dated, and every finding is backed by a verbatim excerpt.